Hipeac GPUs K-means challenge

Optimization Tasks

Please take over as many tasks as you can

Low level

  • (Sakkas) cuSparse on n x k sparse array
  • Point size may not fit in GPU ram
  • Research cuda streams
  • Compute squared distance with cublas
  • Check if cublas blocks or needs Synchronize
  • Make sure all threads are active at all times (Check block and grid size)
  • (Sklikas) Reduce used registers number
  • (Sklikas) Change where centers are saved in memory (Textures, constant)

Report

  • (Tsatiris) First Draft

Evaluation

  • (Kallas) Find dataset with many dimension

Automate evaluation

  • (OK) Write a script that executes all implementations for all datasets
  • (Kallas) Improve the script by keeping the size and dimensionality of each dataset so that we can understand the efficiency of the algorithm for different parametrizations
  • Integrate the script with the gpu implementations of the algorithm

Algorithm

  • (Greg) Simulated Annealing
  • Boruvka
  • Yin Yang Kmeans